A Megakernel per Model A coding-agent framework that writes a single megakernel per model checkpoint per GPU type produced inference engines 1.13–1.77× faster than the fastest general engine (vLLM or SGLang) with one sequence across all eight tested models from 8B to 594B parameters, and up to 2.05× when model storage may change without altering its numerical function. On a production trace with Inkling-2-Small on 4× B200 at 1 to 8 requests per second, 99–100% of requests stayed within service levels versus 82–85% for SGLang, and the longest stall was 0.08–0.15 s versus 2.3–4.4 s. The framework, run with Claude Opus 5.5 and Fable 5.2, gates timing behind a test suite the agent cannot change, requiring numerical agreement with a high-precision reference on unseen inputs during actual execution plus an independent audit; the engines, an OpenAI-API server, the test suite and reproduction tools are released in megakernel-zoo. General inference engines such as vLLM